Ditch the slideshows. Learn how to configure and run OpenMontage—the world's first open-source agentic video production system—to orchestrate Remotion and HyperFrames, build clip-indexed B-roll corpora, and auto-review renders.
Most AI video platforms require a single text prompt and return a simple, hallucinated, 4-second clip. If you want a complete, narrated video, you are usually stuck stitching clips manually or enduring generic "animated PowerPoint slideshows." OpenMontage breaks this paradigm. It is an agentic, multi-stage production orchestra that behaves like a professional editing team. It conducts research, writes narratives, sources authentic B-roll, styles dynamic typography, and reviews its own work.
OpenMontage is the first open-source, pipeline-driven video editor designed for AI agents. By default, it builds a semantic, CLIP-searchable corpus of real motion clips from open-license archives (Archive.org, NASA, Wikimedia, and Pexels) rather than animating still images. It supports both Remotion (React) and HyperFrames (HTML/CSS/GSAP) composition runtimes. You can run it entirely free with offline voice synthesis (Piper TTS) and public archives, or wire up cloud APIs for next-generation cinematic generation.
Traditional code-based video tools utilize rigid programmatic timelines where a script compiles assets at fixed timestamps. OpenMontage is different: it treats video editing as an **agent-in-the-loop choreography**. There is no complex python orchestration code. Instead, your AI coding assistant (Claude Code, Cursor, Codex) reads structured system prompts, interacts with Python tools, and makes creative design decisions guided by Markdown playbooks.
Every production task goes through a standardized seven-stage pipeline:
Research ──> Proposal ──> Scripting ──> Scene Plan ──> Asset Gen ──> Timeline Edit ──> Compile Render
Each stage is governed by a **Director Skill** (Markdown files in skills/pipelines/). The agent reads the skill, executes the appropriate tools, saves checkpoints to a JSON file, and pauses for human approval at crucial creative intersections. This prevents typical AI drift and ensures the project is aligned with your expectations before burning expensive GPU render hours.
A Real B-Roll Corpus: Unlike standard video tools that animate still images using Ken Burns zoom effects, OpenMontage can build a local CLIP-indexed video index. It fetches public domain footage from Archive.org, NASA repositories, and Wikimedia Commons, runs WhisperX and CLIP classification, and cuts authentic historical and documentary footage directly into your timeline.
At proposal time, OpenMontage locks in the render_runtime. It dynamically chooses between two programmatic engines depending on the aesthetic requirements of the brief:
| Feature / Metric | Remotion (React) | HyperFrames (HTML/GSAP) |
|---|---|---|
| Primary Focus | Data-driven explainers, speech sync, avatar presentations | Kinetic typography, SVG animation, product teasers, UI captures |
| Key Animation Model | React spring animations, linear component transitions | Deterministic GSAP timelines, SVG character rigs, shaders |
| Captions Style | Word-by-word active highlighters, TikTok-style pop text | Fluid editorial typography, matte-occlusion composites |
| Scaffold Command | npx create-video-app (Heavy project checkout) |
npx hyperframes init (Lightweight, registry-driven) |
| Best Fit Case | Narrated tutorials with charts and talking-head overlays | Cinematic title loops, website highlight captures, logo stings |
OpenMontage enforces strict runtime isolation. Swapping runtimes mid-pipeline is classified as a governance violation. If your brief calls for rigged character animation, the agent routes the stage to HyperFrames; if it calls for heavy spreadsheet-to-chart visualization with human narration, it targets Remotion.
You can run OpenMontage completely offline without paid API credentials. Below is the quick setup guide for your environment.
Ensure you have the following packages installed on your system:
git clone https://github.com/calesthio/OpenMontage.git cd OpenMontage make setup
Windows Path Fallback: If you are on Windows and the make setup command fails, or if npm install returns an ERR_INVALID_ARG_TYPE, run the command manually:
pip install -r requirements.txt && cd remotion-composer && npx --yes npm install && cd .. && pip install piper-tts && cp .env.example .env
Copy the default environment template and add credentials for the providers you wish to activate. The system falls back cleanly to free local alternatives if fields are left blank.
# --- Media Stocks (Free API Keys) --- PEXELS_API_KEY=your_pexels_key UNSPLASH_ACCESS_KEY=your_unsplash_key # --- AI Media Engines (Optional) --- FAL_KEY=your_fal_ai_key # FLUX images, Kling & Veo video SUNO_API_KEY=your_suno_key # Suno AI musical backing ELEVENLABS_API_KEY=your_elevenlabs_key # Premium TTS synthesis OPENAI_API_KEY=your_openai_key # DALL-E 3 & OpenAI TTS
If you have a dedicated NVIDIA GPU with sufficient VRAM, you can run localized video generation models like Wan 2.1 or Hunyuan entirely locally:
make install-gpu # Add to your .env: VIDEO_GEN_LOCAL_ENABLED=true VIDEO_GEN_LOCAL_MODEL=wan2.1-1.3b # Options: wan2.1-1.3b, cogvideo-5b, ltx2-local
OpenMontage is designed for autonomous agent execution. To prevent agents from running wild and generating huge API bills, it implements strict budget governance:
OBSERVE (track costs only), WARN (log alerts), or CAP (terminate execution on overrun) in your config.json.OpenMontage uses a multi-dimensional selection matrix to automatically choose the best model for each task (e.g. image generation, TTS, or B-roll stock). The selector normalizes task parameters and scores each provider across seven key dimensions:
| Scoring Dimension | Weight | Evaluation Criteria |
|---|---|---|
| Task Fit | 30% | Does the model's output match the target visual style and aspect ratio? |
| Output Quality | 20% | Visual fidelity, resolution, text readability, and absence of distortion. |
| Control Features | 15% | Support for camera motion, negative prompting, and seed controls. |
| Reliability | 15% | API response success rates, request timeout thresholds, and speed stability. |
| Cost Efficiency | 10% | Per-action token or compute pricing rates relative to value. |
| Latency | 5% | Time-to-first-frame (TTFF) or transcription output speeds. |
| Continuity | 5% | Consistency of styles across sequential multi-shot generations. |
Every winning decision and its reasoning logic is recorded inside a cumulative decision_log.json audit trail for review, ensuring complete transparency.
Yes. You can write a custom YAML pipeline manifest inside the pipeline_defs/ folder, define stage boundaries, and link them to custom Markdown instructions in the skills/pipelines/ directory.
The scored selection engine catches credit limits in its reliability checks and automatically falls back to offline providers (like local Piper TTS for audio and B-roll stock video clips from public archives).
OpenMontage runs post-render validation: checks video frames via ffprobe to prevent blank sections, checks audio levels for clipping or prolonged silence, and verifies subtitle synchronizations before displaying the final clip.
Yes. You can route research, scripting, and scene-planning stages to local instances using Ollama or LM Studio by editing the LLM host configuration settings in your system file.